license: apache-2.0 base_model:
- Blackfrost-Research/Muse-Glimmer-30B-Abliterated-BF16 tags:
- muse-glimmer
- abliterated
- mlx
- quantized
- apple-silicon
- 3-bit pipeline_tag: image-text-to-text library_name: mlx
Refusal benchmark — R1-HARMFUL-BENCH-450
Measured on the abliterated model — quantization holds it, no refusal snapback:
| Metric | Result |
|---|---|
| True refusal (harmful, n=300) | 0 / 300 = 0.0% |
| True refusal (full 450) | 0 / 450 = 0.0% |
| Substring-harmful | 0 / 300 |
| Substring-all | 2 / 450 (XSTest false positives) |
| Errors | 0 |
Why this model exists
Muse Glimmer is Meta Superintelligence Labs' 30B agentic, on-device model. This is the abliterated build — refusal behavior removed via a Blackfrost weight-change process — packaged as MLX 3-bit — ~15 GB for Apple-silicon Macs. The local footprint is the product.
Specifications
| Architecture | muse_glimmer — dense, 52 layers, hidden 6656, GQA (32 q / 2 kv), sliding-window attention, + vision tower |
| Base | meta-models/Muse-Glimmer-30B — Meta, Apache-2.0 |
| Transform | Abliterated — refusal behavior removed via a Blackfrost weight-change process; multimodal capability intact |
| Format | MLX 3-bit — ~15 GB |
| Context | 131,072 |
Serving (Apple Silicon / MLX)
pip install mlx-lm
# one-off generate:
mlx_lm.generate --model Blackfrost-Research/Muse-Glimmer-30B-Abliterated-MLX-3bit --prompt "Write a binary search in Python." --max-tokens 1024
# OpenAI-compatible server:
mlx_lm.server --model Blackfrost-Research/Muse-Glimmer-30B-Abliterated-MLX-3bit --port 8080
Or open it directly in LM Studio (MLX runtime) on an Apple-silicon Mac.
Sampling (Meta): temperature 1.0, top_p 0.95, top_k 64. It's a heavy thinker — use a generous max_tokens (≥ 1024) and steer depth with a Reasoning strength: low/medium/high/xhigh system line. Reasoning is returned separately from the final answer.